Short answer: North Small Translate is Cohere’s live, text-only machine-translation model, released on September 10, 2026. Its model ID is north-small-translate-1-0; the Hugging Face repository is CohereLabs/North-Small-Translate-1.0. Cohere documents a sparse mixture-of-experts model with 218 billion total parameters and 25 billion active parameters, 16K input and 16K output limits, and support for 50 languages plus documented locale variants.
There are three different access decisions. You can evaluate it through the Cohere Chat API, where Cohere says it is free until rate limits are reached. You can download the weights from Hugging Face for research and non-commercial use under CC BY-NC 4.0, subject to the Cohere Labs Acceptable Use Policy. For commercial hosted deployment, Cohere documents Model Vault and a commercial license; its current Standard Vault pricing table shows North Small Translate at $17.50 per XL instance-hour, with generative access typically handled through a waitlist. The model is not “small” in local-hardware terms: Cohere’s minimum examples range from one B200 with NVFP4 W4A16 weights to eight H100s for BF16.
Cohere’s quality, throughput and cost figures are vendor-reported. The launch blog says the model scored 83.60 across all languages in Cohere’s WMT26 evaluation, but describes GPT-5.6-Sol as the judge. The official WMT26 task describes human evaluation, so the blog figure should not be treated as an official human-evaluated leaderboard result without the exact submission, test set and protocol. A safe first test should measure your own language directions, terminology, formatting, cost, latency and human review load.
| Decision point | Current primary evidence | Boundary to verify |
|---|---|---|
| Model identity | north-small-translate-1-0; 218B total / 25B active MoE; text in and text out. | Capture the exact model and checkpoint version used by your account or deployment. |
| Language scope | 50 languages in the model card; Cohere documentation lists more than 50 language and locale variants. | Confirm every source-target direction, locale and domain before committing. |
| API evaluation | Chat V2 API; free until rate limits are reached. | North-specific production rate, token price, region and data terms. |
| Open weights | Hugging Face download under CC BY-NC 4.0 for non-commercial use. | Commercial use, derivative terms, acceptable-use obligations and gated repository access. |
| Hosted production | Commercial deployment through Cohere Model Vault; Standard Vault table shows $17.50/XL instance-hour. | Waitlist, tier, replicas, region, monthly/annual commitment, encrypted-vault price and support terms. |
| Local hardware | Minimum examples: 1 B200 at NVFP4 W4A16 or 2 H100s; BF16 requires 4 B200s or 8 H100s. | Actual memory headroom, throughput, power, networking, serving version and workload concurrency. |
What Cohere released
Cohere’s September 10 announcement calls North Small Translate its first translation model in the North family. It is a sparse mixture-of-experts (MoE) model purpose-built for machine translation, rather than a general translation feature hidden inside a broader model. Cohere says the model is available for research and non-commercial use under CC BY-NC 4.0, links to Hugging Face for weights and a hosted demo Space, and points enterprises to RWS’s Language Weaver product for a managed translation and localization offering.
The current Cohere model documentation marks north-small-translate-1-0 as Live and exposes it through the Chat endpoint. The model card identifies the research release as North-Small-Translate-1.0. Keep those names separate in implementation notes: the Cohere API model ID is hyphenated and lower-case, while the Hugging Face repository uses the Cohere Labs organisation and title-case model name.
The model is text-only. Cohere documents text input and generated text output, with a 16K input context and a maximum of 16K output tokens. Those limits are translation-call limits, not a promise that every document can be pasted unchanged: markup, instructions and prompt overhead consume context, and long documents still need segmentation and structure-preserving tests.
Model identity and training lineage
The headline “Small” describes the North family position, not a laptop-sized checkpoint. North Small Translate has 218 billion total parameters and 25 billion active parameters. The Hugging Face model card describes a decoder-only sparse MoE Transformer with 128 experts, eight selected per token and shared experts. It also describes interleaved sliding-window and global attention. These are model-card details, not a guarantee of a particular serving implementation or memory footprint.
A Cohere-authored research paper presents the system as “Cohere CAT+” for the WMT General MT task and says it uses the same foundation as Command A+. The paper describes a non-reasoning base model trained for throughput, followed by a five-step protocol combining coarse and fine-grained supervised fine-tuning, direct preference optimization and online reinforcement learning. It also describes related tasks such as post-editing, quality estimation, structured translation and instruction following.
That paper explains the research lineage but does not grant a separate fine-tuning recipe to customers. I found no current official North Small Translate adapter, LoRA, fine-tuning API or customer-training workflow. Treat customization as an open question for Cohere rather than assuming that an open-weight download can be fine-tuned, commercially redistributed or served with the same quality and support terms.
The four practical access paths
1. Cohere Chat API for a bounded evaluation
Cohere’s North documentation supplies a Chat V2 example using ClientV2 and the model ID north-small-translate-1-0. The docs say North Small Translate is available on the free tier until rate limits are reached. This is useful for a small proof of concept, not an unlimited production grant.
Cohere’s general rate-limit page distinguishes free trial and paid production keys and says trial keys are limited to 1,000 API calls per month. Its Chat table lists several named models at 20 requests per minute but does not list North Small Translate. Do not assign North the rate of Command A Translate or North Mini Code by analogy. Check the current dashboard, response headers and Cohere support for the exact North limit and any production-key requirement.
A source-checked but unexecuted example looks like this:
from cohere import ClientV2
co = ClientV2(api_key="<YOUR_API_KEY>")
response = co.chat(
model="north-small-translate-1-0",
messages=[{
"role": "user",
"content": "Translate everything that follows into French:\n\n"
"Enterprises need accurate translations of business-critical documents."
}],
)
print(response.message.content[0].text)The request pattern follows Cohere’s current model documentation, but it was not executed here because no authorised API credential was available. Before using it in production, add input-size checks, retry/backoff for rate limits, language-direction validation, output parsing and human review for high-risk content. The example does not prove a key, region, model entitlement or billing state.
2. Hugging Face weights for research and non-commercial use
The Cohere Labs model card provides the downloadable open-weight research release and several quantized repositories. Access is gated: the public page says you must agree to the license conditions and share contact information, with collection and sharing governed by the linked Cohere Privacy Policy. A hosted Hugging Face Space lets you try the model before downloading the weights.
The licence is CC BY-NC 4.0, with a Cohere Labs Acceptable Use Policy addendum. “Open weights” therefore does not mean “commercially unrestricted.” A marketing agency, SaaS product or localization vendor should obtain a commercial license before using the weights for paid production work. DMT’s open-weight licensing analysis provides broader context on why a downloadable checkpoint and a commercial right to use it are separate questions. The Hugging Face page also reports that the base repository is not deployed by an Inference Provider, so a one-click hosted provider route should not be assumed.
3. Model Vault for commercial hosted inference
Cohere documents commercial North deployment through Model Vault. A Standard Vault is a Cohere-managed, single-tenant inference environment: Cohere manages serving infrastructure, maintenance, updates, scaling and monitoring while the customer calls a vault endpoint. The North integration can run the North application and control plane in the customer’s VPC while routing model inference to the vault.
Standard Vault pricing currently lists North Small Translate at $17.50 per XL instance-hour. The table shows no L-tier rate for this model. Generative models typically require a waitlist, and Cohere says monthly and annual commitment pricing follows fixed or flex plans and requires contacting Cohere. The hourly figure is a Model Vault instance rate, not a complete translation cost: add the replica range, uptime, autoscaling, storage, network, application, human review and license terms before comparing it with a token-priced API.
Model Vault’s standard deployment is protected in transit and at rest and is logically isolated. If Zero Data Retention (ZDR) is enabled for a Standard Vault standalone deployment, Cohere says it processes inputs and outputs for inference without retaining prompts or responses. Encrypted Vault adds hardware-backed confidential computing and remote attestation, but model and GPU availability differs; verify that North Small Translate is offered in the chosen vault type, tier and region.
4. Private VPC or on-premises deployment
Cohere’s deployment-options material describes private cloud, VPC and on-premises paths for North and says private deployments keep prompts, outputs and fine-tuned models in the customer environment, with Cohere unable to access that data or infrastructure. That is an enterprise deployment option, not an automatic entitlement for every North model. Confirm that North Small Translate, the desired quantization and the selected serving stack are supported for the customer’s cloud, hardware, region, license and support plan.
Language coverage: 50 languages is not a pair-by-pair guarantee
The model card lists 50 languages: Albanian, Arabic, Bengali, Bulgarian, Catalan, Chinese (Simplified), Chinese (Traditional), Croatian, Czech, Danish, Dutch, English, Estonian, Filipino, Finnish, French, German, Greek, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Irish, Italian, Japanese, Korean, Latvian, Lithuanian, Malay, Maltese, Norwegian, Persian, Polish, Portuguese, Romanian, Russian, Serbian, Slovak, Slovenian, Spanish, Swedish, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu and Vietnamese.
Cohere’s current documentation describes more than 50 language and locale variants and calls Modern Standard Arabic, German, French, Japanese, Korean, Russian and Ukrainian its tier-one languages. It separately lists variants such as Egyptian, Modern Standard and Saudi Arabic; Portuguese (Portugal); Norwegian Bokmål; and Serbian (Cyrillic). The launch blog uses “50+ languages,” while the model card uses 50, so use the full current documentation list rather than repeating an unqualified count.
Neither the launch blog nor model card supplies a complete source-target matrix with quality or availability per direction. “Supports 50 languages” should therefore mean “the model is documented for those languages,” not “every one of the 2,500 possible directions has equal quality, locale support or commercial availability.” Test the directions that matter to your business, including regional variants, terminology and script handling.
Hardware, quantization and serving boundaries
| Checkpoint format | Cohere model-card example | What it means for planning |
|---|---|---|
| BF16 | 4 × B200 or 8 × H100 | Highest listed precision; large multi-GPU footprint. |
| FP8 | 2 × B200 or 4 × H100 | Lower-precision checkpoint; confirm serving support and quality on your language pairs. |
| NVFP4 W4A16 | 1 × B200 or 2 × H100 | Smallest listed hardware example; it is still a datacenter-GPU deployment. |
Cohere calls these example minimum requirements. They are not a throughput or total-cost guarantee. A 25B active-parameter MoE can reduce per-token computation, but the 218B total parameter set, temporary buffers, KV cache, networking and concurrency still affect memory and cost. Do not turn “active parameters” into a claim that the model fits in 25GB or on a 25B-class card.
The model card recommends installing a Transformers version from the source repository containing the required model changes and recommends greedy decoding, which Cohere says its production deployment uses. It warns that the older translation pipeline is no longer supported in Transformers 5; loading the model directly or using a compatible earlier Transformers version is required. Generated replies contain structural markers, and Cohere recommends parsing them with its cohere_melody library. The official vLLM example uses tensor parallelism and Cohere’s parser settings, but the example is not a substitute for a compatibility test on the chosen checkpoint.
The following deployment questions remain open in the reviewed sources:
- There is no North-specific fine-tuning or adapter workflow.
- There is no official GGUF or llama.cpp release in the model card. A practitioner discussion says llama.cpp support was not available at launch and speculates about conversion; that is not a supported deployment path.
- There is no model-specific latency, throughput, concurrency, power or networking guarantee for a customer workload.
- There is no complete region-by-quantization-by-vault matrix; Cohere directs customers to the model dropdown or sales contact.
Privacy, data handling and sovereignty
The right privacy answer depends on the path, not simply on the model name. Cohere’s enterprise data commitments say private and third-party deployments do not send customer prompts or generations to Cohere. Its SaaS platform has separate terms, logging and retention controls: enterprise platform inputs and outputs are generally retained for 30 days unless an agreement or approved zero-retention setting changes the treatment, and customers can opt out of model training in dashboard data controls.
For a trial or research key, do not paste personal or confidential data merely because the request is free. Cohere’s privacy policy says the platform is not intended to process personal information for trial users and researchers, and points enterprise users toward a commercial agreement and data-processing addendum where appropriate. Verify the data path, sub-processors, retention, training opt-out and incident process with your legal and security teams.
For a managed private deployment, distinguish Standard Vault from Encrypted Vault. Standard Vault offers logical isolation and protection in transit and at rest. Encrypted Vault adds protection in use within hardware-backed trusted execution environments and lets the client verify remote attestation before sending data. Neither model name nor CC licence alone answers data-residency, deletion, access-log or cross-border questions; the contract and deployment configuration do.
Pricing: free to try is not free at scale
Cohere’s model page says North Small Translate is free for trial and production keys until rate limits are reached. The general pricing documentation says trial keys are free but limited, while production usage is metered. The exact North token price was not published in the current rate-limit table reviewed for this article, so do not invent an input/output rate from Command A Translate or another model.
For commercial hosted inference, the current Standard Vault table gives one concrete anchor: $17.50 per XL instance-hour for North Small Translate. Cohere says generative model access typically involves a waitlist and that fixed/flex monthly or annual plans depend on the model and performance tier. Encrypted Vault has separate pricing. A buyer should request a quote that states model version, tier, replica minimum and maximum, autoscaling, pause billing, region, support, SLA, commercial licence, ZDR and any North application costs.
Cohere’s launch blog also reports a cost-per-task comparison: an 80.1 score at $0.000676 per task using 661 tokens on average, compared with $0.038928 for Gemini 3.1 Pro Preview (high). That calculation uses provider prices and an evaluation task, not a full ownership model. It excludes or may not expose hardware, Model Vault capacity, engineering, data preparation, human post-editing, retries, region, contract minimums and the cost of quality failures. Use it as a vendor scenario, not as a savings forecast. DMT’s API access and pricing checklist is a useful adjacent reminder to record the exact model ID, billing basis and rate limits for any comparison.
What the evaluation claims do—and do not—show
Cohere’s launch blog reports an 83.60 WMT26 all-languages score for North Small Translate and 84.36 for an “Agentic” multi-pass workflow. It compares North with Qwen, GLM, DeepL, Gemma, Google Translate and other models. The blog says the graphs were benchmarked using GPT-5.6-Sol as a judge. It also reports regional averages, throughput tests against Gemma 4 31B TP1, a long-context score of 48.9, and the cost-per-task calculation above.
These are useful vendor disclosures, but they need careful labels:
| Claim or test | Reported setup | Limit before using it as a buying conclusion |
|---|---|---|
| WMT26 quality | 83.60 all languages; 84.36 for the optional Agentic workflow; blog graphs use GPT-5.6-Sol as judge. | Confirm testset, prompt, language-pair weighting, judge calibration and whether the number is an official WMT result. |
| Official WMT context | WMT26’s General MT task uses unseen documents across languages/domains and human evaluation. | Cohere’s blog does not publish a North submission ID or human-evaluation report connecting its score to the official result. |
| Long context | Cohere says it translates two book chapters in one call and scores paragraph quality with xCOMET-XL. | Test your document structure, segmentation, terminology and failure recovery; a long-context score is not a guarantee for every document. |
| Throughput | Internal identical-hardware/concurrency tests report 112 vs. 81 output tokens per second at low concurrency and 39 vs. 30 at high concurrency against Gemma 4 31B TP1. | Not an end-to-end latency SLA or a promise across quantizations, hardware, batch size or network. |
| Cost per task | Vendor calculation using provider prices and average tokens. | Not total cost of ownership, commercial licence price or a forecast for your workload. |
The official WMT26 General MT task says its goal is broad evaluation across languages, domains and genres and that systems are evaluated with humans. The Cohere-authored North Small Translate paper adds research details on difficulty sampling, SFT, DPO and online reinforcement learning, and reports additional evaluation categories such as terminology, structured translation and long context. Neither source is an independent commercial deployment study. No independent North Small Translate production benchmark or customer quality report was accessed for this package.
Which deployment fits which translation job?
| Need | Best starting path | Why and what to check |
|---|---|---|
| Small non-sensitive prototype | Cohere Chat API or Hugging Face Space | Fastest discovery; free/rate-limited API or demo. Do not use confidential data or infer production limits. |
| Research or non-commercial self-hosting | Hugging Face NVFP4, FP8 or BF16 checkpoint | Maximum model control, but large GPU footprint and CC BY-NC/acceptable-use obligations. |
| Commercial managed inference | Model Vault Standard | Dedicated Cohere-managed serving and a published XL hourly anchor; waitlist, tier, region and commercial licence remain. |
| Highly sensitive regulated workload | Model Vault Encrypted or approved private deployment | Encrypted Vault adds attestation; VPC/on-prem options add customer control. Verify North model/GPU availability and contract. |
| Edge, CPU or ordinary workstation | Do not assume fit | The smallest Cohere example is still one B200 or two H100s; compare a smaller translation model instead if hardware is fixed. |
A safe first evaluation for marketing and localization teams
- Freeze the job. Choose one workflow, such as translating product pages, support macros or internal campaign briefs. Do not mix legal, customer-service and creative copy into one score.
- Build a representative test set. Include the exact source-target directions, regional variants, terminology, HTML/JSON/CSV structures, long documents, numbers, brand names and opt-out content. Use approved or synthetic data for the first pass.
- Record the route. Capture API model ID or Hugging Face repository/commit, quantization, Transformers/vLLM version, prompt, decoding settings, context limit and deployment region. A model family name is not enough.
- Compare against an incumbent. Run the current translation system, North Small Translate and a human reference or post-edit sample under the same input and output rules. Keep the evaluator blind where practical.
- Score more than fluency. Track adequacy, omissions, additions, terminology accuracy, numbers, names, formatting, locale, tone, safety and reviewer edits. For marketing pages, check whether links, markup, offers and regulated wording survive translation.
- Measure operating cost. Record input/output tokens, retries, rate-limit events, queue time, GPU hours or Model Vault replicas, human post-edit time and failed translations. Do not use the blog’s per-task figure as your ledger.
- Test limits and failure modes. Deliberately approach 16K input/output boundaries, malformed markup, mixed languages, unsupported locales, long documents and concurrent requests. Add truncation and human-escalation paths.
- Gate publication or customer delivery. Keep the first outputs in review. Expand only when the quality threshold, cost ceiling, data contract, licence and rollback plan are approved by the content and security owners.
This sequence is a practical evaluation recommendation, not a Cohere product requirement. No API key, Hugging Face download, GPU deployment or live translation call was executed in this package.
What it changes for digital marketing—and what it does not
North Small Translate could be relevant when a team needs to localize product content, campaign variants, help-centre material, customer communications or internal documentation across the documented languages. Open weights can offer more control over hosting and integration than a fixed translation SaaS, while Model Vault can reduce the serving burden for a commercial deployment. Those are deployment trade-offs, not evidence that every translated asset will rank, convert or satisfy local audiences. Keep the model decision separate from platform targeting: DMT’s Google Ads language-targeting guide covers campaign-language controls, while this page covers translation-engine access and deployment.
Keep human ownership for claims, offers, consent language, pricing, legal text and culturally sensitive copy. A translation benchmark can reward adequacy while missing a brand-specific term, a prohibited claim or a market-specific nuance. For SEO, preserve structured markup, canonical URLs, hreflang decisions and editorial review as separate controls; a translation model does not automatically create a sound international search architecture.
The appropriate buying question is therefore not “is North Small Translate the best translator?” It is “for which language directions and document types does its quality, hardware path, licence and data boundary beat our current workflow?”
Verified facts, vendor claims and unknowns
| Type | Current statement |
|---|---|
| Verified in current Cohere documentation | Live model ID north-small-translate-1-0; 218B total / 25B active; 16K input and 16K output; text-only; Chat endpoint; more than 50 language/locale variants. |
| Verified in the model card | Research weights, CC BY-NC 4.0 plus acceptable-use obligations, three quantization families and example GPU requirements. |
| Vendor-reported performance | WMT26 scores, regional comparisons, throughput, long-context and cost-per-task figures in Cohere’s launch material. |
| Not established | All-pair quality, universal API rate, token price, full region matrix, fine-tuning support, GGUF/llama.cpp support, independent customer outcomes and a production SLA. |
Frequently asked questions
Is North Small Translate open source?
Cohere calls it an open-weight research release, but the Hugging Face model card applies CC BY-NC 4.0 and the Cohere Labs Acceptable Use Policy. That permits research and non-commercial use under the stated terms; it is not a blanket commercial licence.
What is the North Small Translate API model ID?
The current Cohere model ID is north-small-translate-1-0. The Hugging Face repository is CohereLabs/North-Small-Translate-1.0. Record the exact path and version because API and self-hosted checkpoints are different access routes.
Can I use it for free in production?
Cohere says trial and production keys are free until rate limits are reached, but production use is still governed by key, pricing, terms and rate-limit rules. The open weights are non-commercial. A paid product or service needs a commercial licence and a confirmed production route such as Model Vault.
Does it support every language pair equally?
No such guarantee is published. Cohere documents 50 languages and additional locale variants, but the reviewed pages do not provide a quality-ranked pair-by-pair matrix. Test your exact source-target directions and locales.
Can it run on one GPU?
The smallest listed example is one B200 with NVFP4 W4A16 weights. Cohere also lists two H100s for that quantization. Those are example minimums, not a promise of your required memory, throughput or concurrency. BF16 requires four B200s or eight H100s in the model card’s examples.
Is the 83.60 WMT26 score independently verified?
Not by the evidence reviewed here. Cohere reports the score and says its graph used GPT-5.6-Sol as a judge, while the official WMT26 task describes human evaluation. Treat the score as a Cohere-reported evaluation until the exact submission and human-evaluation evidence are available.
Bottom line
North Small Translate is a real, current Cohere release with a clear model ID, open-weight research path, Cohere API evaluation path and commercial Model Vault path. Its 50-language scope, 16K input/output window and quantized checkpoints make it worth testing for controlled localization and documentation jobs. Its 218B total-parameter footprint means “small” should not be read as local or inexpensive by default.
Before adopting it, verify the licence, source-target directions, API rate and price, deployment region, checkpoint, hardware, data retention, fine-tuning expectations and human-review cost. Treat WMT, throughput and cost numbers as vendor evidence, then run a fixed, representative evaluation against your incumbent. That is the difference between a promising translation release and a safe production decision.
Sources checked
- Cohere: Introducing North Small Translate, September 10, 2026.
- Cohere documentation: North Small Translate, for model ID, limits, languages, API, licensing and deployment.
- Cohere Labs model card on Hugging Face, for open-weight terms, quantizations, hardware examples and serving caveats.
- Cohere model overview, for Live status and North API/Model Vault positioning.
- Cohere API keys and rate limits, for trial/production key boundaries and the absence of a North-specific row in the current table.
- Cohere Standard Vault pricing, for the $17.50 XL instance-hour listing and waitlist caveat.
- Cohere Standard Vault overview and Model Vault with North, for isolation, ZDR and North routing.
- Cohere Enterprise Data Commitments and Cohere Privacy Policy, for SaaS/private-deployment data boundaries and retention context.
- Cohere deployment options, for VPC, on-premises and Model Vault availability positioning.
- North Small Translate: Advanced Cost-Effective Translation (Cohere CAT+), submitted September 12, 2026, for research training and evaluation context.
- WMT26 General Machine Translation task, for the official shared-task scope and human-evaluation boundary.
Scope note: This article reports Cohere, Cohere Labs, Hugging Face, WMT26 and Cohere-authored research material accessed on September 16, 2026. It attributes vendor performance and cost claims, keeps API, research-weight and Model Vault access separate, and does not claim universal language-pair quality, commercial weight rights, exact North API pricing, complete region coverage, fine-tuning support, independent outcomes or a production SLA. The code example and deployment guidance were source-checked but not executed with an authorised credential or GPU environment.
